comparison of the estimation capabilities of response surface

7/30/2019 Comparison of the Estimation Capabilities of Response Surface

1/12

O R I G I N A L P A P E R

Comparison of the estimation capabilities of response surfacemethodology and artificial neural network for the optimization

of recombinant lipase production by E. coli BL21

Rubina Nelofer Ramakrishnan Nagasundara Ramanan

Raja Noor Zaliha Raja Abd Rahman

Mahiran Basri Arbakariya B. Ariff

Received: 14 February 2011 / Accepted: 9 July 2011 / Published online: 11 August 2011

Society for Industrial Microbiology 2011

Abstract Response surface methodology (RSM) and

artificial neural network (ANN) were used to optimize theeffect of four independent variables, viz. glucose, sodium

chloride (NaCl), temperature and induction time, on lipase

production by a recombinant Escherichia coli BL21. The

optimization and prediction capabilities of RSM and ANN

were then compared. RSM predicted the dependent vari-

able with a good coefficient of correlation determination

(R2) and adjusted R2 values for the model. Although the R2

value showed a good fit, absolute average deviation (AAD)

and root mean square error (RMSE) values did not support

the accuracy of the model and this was due to the inferi-

ority in predicting the values towards the edges of the

design points. On the other hand, ANN-predicted values

were closer to the observed values with better R2, adjusted

R2, AAD and RMSE values and this was due to the

capability of predicting the values throughout the selected

range of the design points. Similar to RSM, ANN could

also be used to rank the effect of variables. However, ANN

could not predict the interactive effect between the vari-ables as performed by RSM. The optimum levels for glu-

cose, NaCl, temperature and induction time predicted by

RSM are 32 g/L, 5 g/L, 32C and 2.12 h, and those by

ANN are 25 g/L, 3 g/L, 30C and 2 h, respectively. The

ANN-predicted optimal levels gave higher lipase activity

(55.8 IU/mL) as compared to RSM-predicted levels

(50.2 IU/mL) and the predicted lipase activity was also

closer to the observed data at these levels, suggesting that

ANN is a better optimization method than RSM for lipase

production by the recombinant strain.

Keywords Process optimization Response surfacemethodology Artificial neural network Lipase production Recombinant Escherichia coli

Introduction

Optimization of medium components and environmental

factors is an important step in a high-performance fermen-

tation process [11]. Various techniques have been proposed

and used for the optimization of industrial processes. The

classical technique of one parameter at a time is time con-

suming and not effective for the identification of interac-

tions and also for the predictions of parameters involved in

the process. Nowadays, mathematical methods such as

response surface methodology (RSM) [1, 11, 20, 23] and

artificial neural network (ANN) [9, 13, 27] are commonly

used for modelling and optimization of processes.

RSM is based on a collection of statistical and mathe-

matical techniques. This technique is useful in developing,

improving and optimizing processes in which a response of

interest is influenced by several independent variables and

R. Nelofer A. B. Ariff (&)Department of Bioprocess Technology,

Faculty of Biotechnology and Biomolecular Sciences,

University Putra Malaysia, 43400 Serdang, Selangor, Malaysia

e-mail: [email protected]

R. N. RamananChemical and Sustainable Process Engineering Research Group,

School of Engineering, Monash University,

46150 Bandar Sunway, Selangor, Malaysia

R. N. Z. R. A. Rahman

Department of Microbiology, Faculty of Biotechnology

and Biomolecular Sciences, University Putra Malaysia,

43400 Serdang, Selangor, Malaysia

M. Basri

Department of Chemistry, Faculty of Science,

University Putra Malaysia, 43400 Serdang, Selangor, Malaysia

123

J Ind Microbiol Biotechnol (2012) 39:243254

DOI 10.1007/s10295-011-1019-3


2/12

the objective is to optimize this response. Besides analys-

ing the effects of the independent variables, this experi-

mental methodology generates a mathematical model,

which can be used to describe the process for better

understanding. Before applying the RSM, some pre-

liminary studies are needed such as selection of the

experimental design [5]. If the number of variables is too

large, screening of the significant variables shall be carriedout first, prior to selection of appropriate optimization

design. Various experimental designs may be used for

RSM optimization [7, 8, 25]. RSM has many advantages

over the conventional one parameter at a time technique,

but it is not applicable to all optimization and modelling

studies. The major drawback of RSM is the need for a

second-order polynomial to fit the data [3]. All systems

containing curvature such as symmetrical or non-symmet-

rical bell-shaped curves may not be well explained by the

second-order polynomial [3, 6].

ANNs are computational models formed from hundreds

of single units, artificial neurons, inspired by biologicalneurons and connected with coefficients (weights) which

constitute the neural structure [7]. These neurons are

sometimes called processing elements (PE) as they process

information. These weights are just like the synaptic

activity in a biological neuron. The weights of the inputs

are summed, and the threshold subtracted, to determine the

activation of the neuron [22]. The other important capa-

bility of neural networks is that they can learn the input/

output relationship through training. ANN analysis is quite

flexible as regards to the amount and form of the training

(experimental) data, which makes it possible to use more

informal experimental designs than with statistical

approaches [22].

A neural network does not need any model or screening

before the development of a network. Neural networks may

be applied on designed data or on the data that is not sta-

tistically designed. Sufficient data with all possible oper-

ating conditions of input variables are needed to develop a

neural network. A network model is then constructed

according to the systems behaviour. The constructed model

may be used for predictions and other applications within

the assessed operating conditions. Since the regression

analysis is dependent on predetermined statistically sig-

nificant levels, the less significant factors are not included

in the model. ANN uses all the data making the model

more accurate [7]. A neural network can perform tasks that

cannot be performed by linear programming. If an element

of the neural network fails, the network can still continue to

perform the task owing to its parallel nature. Neural net-

works learn and there is no need for reprogramming [15].

The main disadvantage of ANN is the requirement of

training in order to operate. A neural network needs to be

emulated because the architecture is different from that of

microprocessors [15], where high processing time is

required for a large network. Different architectures may

also be involved in ANN which requires different types of

algorithms.

The development of accurate models for a biological

reaction on chemical and physical bases is still a critical

challenge, mainly due to the non-linear nature of the bio-

chemical network interactions. The use of advanced non-linear data analysis techniques such as ANN has been

applied in various areas such as food science [26], bio-

technology [12], chemical processes [4], equipment

development [22] and biochemical engineering [13].

Comparative studies of ANN and RSM for fermentation

processes employing wild strains have been reported [79].

To our knowledge, comparison of ANN and RSM for

optimization of cultural conditions for enzyme production

by a recombinant strain has not been reported in the

literature.

The objective of the present study was to compare the

efficiency of ANN in modelling and optimization oflipase production by a recombinant E. coli BL21 with

optimization using RSM from data of the previous study

[21]. The optimization and prediction capabilities of RSM

and ANN were compared by the coefficient of correlation

determination (R2), adjusted R2, absolute average devia-

tion (AAD) and root mean square error (RMSE) values

for the models.

Materials and methods

Microorganism and inoculum preparation

The microorganism used in this study was E. coli BL21

(DE3) pLysS [9, 14] harbouring the organic solvent toler-

ant and thermostable lipase gene of Bacillus sp. 42 [10].

Inoculum was prepared by adding a single colony grown

from an LB agar plate in 50 mL LB broth in 250-mL

screw-cap Schott Duran bottles and incubating in a rotary

shaker with a shaking speed of 200 rpm at 37C for

1618 h.

Lipase production

All fermentations were conducted under aerobic conditions

in 250-mL screw cap Schott Duran bottles with 50 mL

production medium (5 g/L yeast extract, 10 g/L tryptone,

19 g/L NaCl and 1090 g/L glucose). The pH of all

media was adjusted to 7 using either 0.1 M HCl or NaOH

prior to sterilization. Ampicillin (50 lg/mL) and chlor-

amphenicol (35 lg/mL) were added to all media to inhibit

growth of bacteria without lipase gene. Fermentation was

carried out in a rotary shaker agitated at 200 rpm for 24 h.

244 J Ind Microbiol Biotechnol (2012) 39:243254

123


3/12

The culture pH was not controlled throughout the culti-

vations but the pH was measured at time intervals. In all

fermentations, isopropyl b-D-1-thiogalactopyranoside

(IPTG) at a concentration of 0.5 mM was used as an

inducer. Induction time was varied from 1 to 5 h and the

fermentation was carried out for 24 h. The variations in

medium components and other fermentation variables are

given in Table 1.

Experimental design

The optimization of fermentation conditions was con-

ducted using the PlackettBurman (PB) design as described

earlier [21]. In this experimental design, four significant

variables (glucose, NaCl, temperature and induction time)

were selected. A total of 32 experiments were conducted

according to Box-Wilson (BW) 24 full factorial central

Table 1 Box-Wilson 24 factorial central composite design for optimization of lipase production by recombinant E. coli used for RSM [21] and

ANN

Exp. no. Glucose

(X1) (g/L)

NaCl (X2)

(g/L)

Temperature

(X3) (C)

Induction

(X4) time (h)

Lipase activity (Y) (IU/mL) Final

culture

pHObserved Predicted by

RSM (% differencea)

Predicted by

ANN (% differencea)

1 70 7 43 4 04.2 07.2 (69) 04.9 (15) 4.93

2 70 7 43 2 12.9 14.4 (10) 12.9 (0.5) 4.91

3 70 7 31 4 18.9 18.2 (4) 18.8 (0.6) 4.52

4 70 7 31 2 29.7 30.0 (1) 29.2 (1) 4.53

5 70 3 43 4 06.0 08.0 (33) 06.6 (9) 4.95

6 70 3 43 2 15.7 16.9 (8) 15.9 (2) 4.94

7 70 3 31 4 22.0 20.6 (6) 21.5 (2) 4.67

8 70 3 31 2 35.3 34.2 (3) 35.6 (0.8) 4.69

9 30 7 43 4 0.02 1.9 (9076) 0.01 (33) 5.29

10 30 7 43 2 12.9 14.7 (13) 12.4 (4) 5.29

11 30 7 31 4 22.6 21.7 (4) 22.3 (2) 5.28

12 30 7 31 2 40.3 39.0 (3) 42.3 (5) 5.29

13 30 3 43 4 0.04 -0.02 (134) 0.03 (31) 5.30

14 30 3 43 2 12.9 14.4 (11) 12.6 (3) 5.30

15 30 3 31 4 21.9 21.3 (3) 21.6(2) 5.29

16 30 3 31 2 43.1 40.4 (6) 44.0 (2) 5.30

17 90 5 37 3 23.4 21.6 (8) 23.7 (2) 4.21

18 10 5 37 3 21.8 22.5 (3) 21.9 (0.3) 7.27

19 50 9 37 3 26.9 24.7 (8) 26.6 (1) 4.28

20 50 1 37 3 25.9 26.9 (4) 25.6 (1) 4.45

21 50 5 49 3 0.05 -5.7 (11132) 0.04 (15) 5.74

22 50 5 25 3 26.6 31.3 (17) 26.9 (1) 4.78

23 50 5 37 5 02.9 01.9 (33) 02.7 (10) 4.28

24 50 5 37 1 28.4 28.3 (0.4) 28.6 (0.4) 4.29

25 50 5 37 3 34.2 35.3 (3) 34.7 (2) 4.30

26 50 5 37 3 35.8 35.3 (1) 34.7 (3) 4.3127 50 5 37 3 34.9 35.3 (1) 34.7 (0.6) 4.31

28 50 5 37 3 35.3 35.3 (0.2) 34.7 (1) 4.30

29 50 5 37 3 34.9 35.3 (1) 34.7 (0.6) 4.31

30 50 5 37 3 35.5 35.3 (0.6) 34.7 (2) 4.29

31 50 5 37 3 36.0 35.3 (2) 34.7 (3) 4.30

32 50 5 37 3 35.9 35.3 (2) 34.7 (3) 4.29

The italic, bold and normal values represent the experiments used for selection, training and testing, respectively, by the selected ANNa % difference was calculated as the % difference between the observed value and corresponding predicted value over the observed value

J Ind Microbiol Biotechnol (2012) 39:243254 245

123


4/12

composite design (CCD). Each variable was set at five

different levels of variations (Table 1). The first 16

experiments (24 = 16, factorial CCD) were at factorial

points, eight at axial points (a = 2) and eight replications

for the central points.

Response surface methodology

The optimization results of lipase fermentation conditions

by recombinant E. coli using RSM as reported in the pre-

vious study [21] with some extensions in experiments and

statistical analysis were used in this study. In the RSM

method, a second-order model (Eq. 1) was used to calcu-

late the predicted response and optimal levels:

Y b0 b1X1 b2X2 b3X3 b4X4 b11X21

b22X22 b33X

23 b44X

24 b12X1 X2 b13X1

X3 b14X1 X4 b23X2 X3 b24X2

X4 b34X3 X4 1

where Y represents the response variable and b0 is the

interception coefficient. b1, b2, b3 and b4 are coefficients of

the linear effects, b11, b22, b33 and b44 are coefficients of

quadratic effects and b12, b13, b14, b23, b24 and b34 are

coefficients of interaction effects for the four independent

variables (X1 = glucose, X2 = NaCl, X3 = temperature

and X4 = induction time).

Artificial neural network

For comparison, the same data of lipase fermentation by

recombinant E. coli applied for optimization using RSM, asreported in the previous study [21], were applied to ANN.

The intelligent problem solver in STATISTICA software

version 7 was used to construct the regression-based net-

works from the data. A total of 60 different trained net-

works were observed for selection on the bases of the

highest coefficient of correlation determination (R2) and

the lowest selection error. A multilayer perception network

(MPN) was selected from them. Back propagation (BP)

and conjugate gradient descent (CG) algorithms were used

in the training of the neural network on the basis of varying

input/output pair data sets. The experiments used for

selection (8), training (16) and testing (8) are indicated inTable 1. The topology of the network consists of three

layers with one hidden layer.

Comparison of optimization capability of ANN

and RSM

Adjusted R2, AAD and RMSE were calculated in addition

to the R2 for the comparison of estimation capabilities of

RSM and ANN. The R2 was calculated using Eq. 2:

R2

Pi1n Xi Yi

2Pi1n

"Yi Yi2

2

where X is the predicted lipase activity (by either RSM or

ANN), Yis the observed lipase activity and "Y is the average

observed lipase activity.

The adjusted R2 was calculated using Eq. 3:

AdjustedR2 1 1 R2 N 1

N K 1

!3

where N is the total number of observations and K is the

number of input variables.

The AAD was calculated using Eq. 4:

AAD Xp

i1jyi;exp yi;calj=yi;exp

h i.P

n o 100 4

where yi,exp and yi,cal are the experimental and calculated

responses, respectively, and P is the number of

experiments.

The RMSE was calculated using Eq. 5:

RMSE

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiPyi;exp yi;cal

2

n

s5

where yi,exp is the experimental response, yi,cal is the cal-

culated response and n is the number of experiments.

Statistical analysis

The statistical analysis of data and plots were constructed

using STATISTICA software version 7. The analysis of

variance (ANOVA) was employed to determine the sig-

nificance of model parameters in RSM. R2 and adjusted R2

values were calculated to evaluate the performance of the

regression model. The optimum levels of the selected

variables were obtained from the desirability charts (data

not shown). The intelligent problem solver in STATISTICA

neural network was used to construct various neural

networks, the best of which was selected and used for

prediction and optimization. Sensitivity analysis was con-

ducted by the selected ANN to rank the input variables.

Analytical procedures

Culture samples were centrifuged at 10,7009g (TA-14-50,

Allegra250R, Beckman Coulter, USA) for 10 min to

obtain the cell pellets. The collected cell pellets were

washed and resuspended in 20 mM phosphate buffer at pH

7 and then lysed by sonication for 2 min on ice. Superna-

tant was used for lipase determination after removal of the

cell debris by centrifugation. Lipase activity was deter-

mined according to the method proposed by Hamid et al.

[14], which is a modified form of Kwon and Rhees [17]

method. In this method, 1 mL of diluted sample was mixed


123


5/12

with 2.5 mL olive oil emulsion in phosphate buffer with a

ratio of 1:1. Subsequently, 20 lL of 20 mM CaCl2 was

added to the mixture. The reaction was carried out at 60C

in a water bath, agitated at 200 rpm for 30 min. The

reaction was stopped by the addition of 6 N HCl (1 mL).

The free fatty acids liberated by the action of lipase were

extracted in 5 mL isooctane. Pyridine cupric acetate

reagent (1 mL) was then added to the extracted free fattyacids in isooctane (4 mL) and vortexed. The absorbance of

the upper layer was read at 715 nm. A standard curve

constructed from different concentrations of oleic acid was

used to determine the concentration of free fatty acids in

each sample. One unit of lipase activity is defined as 1 lM

free fatty acids released per minute. The cell concentration

was determined as dry cell weight by drying the cell pellet

at 80C until a constant weight was achieved, normally for

at least 24 h.

Results

Optimization using RSM

The full quadratic second-order model obtained by multi-

ple regression analysis of the experimental data by apply-

ing RSM was expressed in Eq. 6. This model was used for

the prediction of lipase activity.

Y 110:933 0:024X1 4:657X2 8:401X3

11:996X4 0:008X21 0:593X

22 0:157X

23

5:046X24 0:017X1 X2 0:018X1 X3

0:033X1 X4 0:069X2 X3 0:221X2 X4

0:193X3 X4 6

where Y is the lipase activity and X1, X2, X3 and X4are glucose, NaCl, temperature and induction time,

respectively.

The model is a highly significant model according to the

statistical analysis (Table 2). The calculated F value was

16.4 with a very small P value (0.0008). The high R2 value

(0.97) supported the models accuracy. Only 3% of the

total variations were not explained by the model, as is

obvious from the R2 value. The significance of the model is

also represented by the value of the adjusted R2 (0.96).

The regression coefficient values and P values (Table 2)

indicated that the main effects of temperature and induction

time were significant. Quadratic effects of all four variables

glucose, NaCl, temperature and induction time were sig-

nificant. The maximum effect was due to the temperature

followed by induction time, NaCl and glucose sequentially.

The temperatureglucose and induction timeglucose

interaction effects were significant. The interaction effects

are represented by 3D surface plot (Figs. 1, 2).

Optimization using ANN

The selected network (MPN) has a better R2 (0.999) than

RSM. Predicted values of this model were also closer to the

observed values than the RSM-predicted values (Table 1).

The topology of the network consisted of three layers

(4:9:1), an input layer consisting of four fermentation

variables, a middle hidden layer of nine neurons and oneoutput layer for lipase activity. The activation level of the

neurons for ANN processing is represented by different

colours (Fig. 3). The optimum levels of glucose (25 g/L),

NaCl (3 g/L), induction time (2 h) and temperature (30C)

predicted by ANN were different from those predicted by

RSM. The highest effect was from the temperature, fol-

lowed by induction time as calculated by the sensitivity

analysis (Table 3). These results are in agreement with

those obtained by RSM. Glucose and NaCl represented the

third and fourth in terms of significant effect, respectively.

The interaction effects are represented by surface plots

(Figs. 4, 5).

Comparison of optimization using RSM and ANN

The predicted values using the selected model of RSM and

selected network of ANN for the experimental runs with

their percentage differences from the observed lipase val-

ues are provided in Table 2. In most of the cases ANN-

predicted values were closer to the observed values. This

difference between the RSM and ANN was more promi-

nent in the experiments at the edge points such as in

experiments 1, 2, 5, 9, 10, 13, 14 and 21, whereas at the

centre points the predicted levels by RSM and ANN were

more similar.

Verification experiments including RSM- and ANN-

predicted optimum levels for the tested variables are rep-

resented in Table 4. The RSM-predicted levels were

32.4 g/L, 5 g/L, 31.7C and 2.121 h for glucose, NaCl,

temperature and induction time, respectively. The pre-

dicted lipase activity at these optimum levels was 48.9 IU/

mL. The ANN-predicted levels were 25 g/L, 3 g/L, 30C

and 2 h for glucose, NaCl, temperature and induction time,

respectively, with 56.3 IU/mL predicted activity. Experi-

ments were conducted in triplicate at these optimum levels

to calculate the observed response. The observed lipase

activity at the predicted optimum levels of tested variables

was 50.2 IU/mL at RSM optimum levels and 55.8 IU/mL

at ANN predicted levels (Table 4). The lipase activity

predicted by ANN was closer to the observed lipase

activity at mean levels of four variables, at the levels used

before optimization and in some other verification experi-

ments near ANN-predicted optimum levels as compared to

RSM (Table 4). In these verification experiments only

glucose concentration was varied because other parameters


123


6/12

Table 2 Analysis of variance for optimization of lipase production by recombinant E. coli using Box-Wilson design (data shown in Table 1),

calculated by RSM regression [21]

Variables Analysis of variance Parameter estimates

SS Deg. of

freedom

MS F Estimates t values P values Confidence limits

-95% ?95%

Intercept 103.5437 1 103.5437 16.4015 2110.933 24.0499 0.000832 2168.724 253.1414

Glucose (X1) 0.0560 1 0.0560 0.0089 0.024 0.0942 0.926090 -0.520 0.5686

Glucose2 (X12) 325.1179 1 325.1179 51.4991 20.008 27.1763 0.000002 20.011 20.0059

NaCl (X2) 20.5663 1 20.5663 3.2577 4.657 1.8049 0.088831 -0.787 10.1004

NaCl2 (X22) 166.1063 1 166.1063 26.3114 20.593 25.1295 0.000084 20.837 20.3490

Temperature (X3) 388.8201 1 388.8201 61.5896 8.401 7.8479 0.000000 6.143 10.6599

Temperature2 (X32) 939.7712 1 939.7712 148.8609 20.157 212.2009 0.000000 20.184 20.1296

Induction time (X4) 32.5716 1 32.5716 5.1594 11.996 2.2714 0.036401 0.854 23.1393

Induction time2 (X42) 752.0353 1 752.0353 119.1233 25.046 210.9144 0.000000 26.021 24.0704

Glucose (X1) * NaCl (X2) 7.8338 1 7.8338 1.2409 -0.017 -1.1140 0.280802 -0.051 0.0156

Glucose (X1) * temperature

(X3)

76.2758 1 76.2758 12.0822 0.018 3.4759 0.002891 0.007 0.0292

Glucose (X1

) * induction time

(X4)

2.4380 1 2.4380 0.3862 0.033 0.6214 0.542560 -0.078 0.1430

NaCl (X2) * temperature (X3) 30.4991 1 30.4991 4.8311 0.069 2.1980 0.042092 0.003 0.1353

NaCl (X2) * temperature (X3) 3.1237 1 3.1237 0.4948 0.221 0.7034 0.491317 -0.442 0.8836

Temperature (X3) * induction

time (X4)

21.4360 1 21.4360 3.3955 0.193 1.8427 0.082885 -0.028 0.4138

Error 107.3224 17 6.3131

Bold letters represent the significant variables and their calculated values. SS and MSare sum of squares and mean sum of squares, respectively

The asterisk between two variables represents interaction effects of the two variables

Fig. 1 Surface plot obtainedfrom optimization using RSM

for the combination effect of

temperature and glucose on

lipase production by

recombinant E. coli by keeping

other parameters constant [21]


123


7/12

are not much different in ANN- and RSM-predicted

optima. There is only a small difference between the

observed and predicted response. Four evaluation param-

eters, namely R2, adjusted R2, AAD and RMSE, were used

to compare RSM and ANN (Table 5). The selected ANN

has higher values of R2 (0.999) and adjusted R2 (0.988)

than those obtained by RSM (0.97 and 0.965, respectively).

On the other hand, the values of AAD (5.09) and RMSE

(0.63) for ANN were lower than those obtained by RSM

(26.38 and 1.48, respectively). These results indicate that

ANN was superior to RSM for the optimization of lipase

production by recombinant E. coli.

Discussion

Optimization is used to increase the performance of a

system. Traditionally, the one parameter at a time tech-

nique is used to optimize bioprocesses. RSM has many

advantages compared to the classical one variable at a time

optimization technique. ANN has received increasing

interest over the last few years, and has been successfully

applied across an extraordinary range of problem domains,

in areas as diverse as business, medicine, engineering,

geology, biotechnology, bioprocessing and physics.

The second-order model obtained from RSM for opti-

mization of recombinant lipase production was verified by

ANOVA and R2. In the ANOVA test, calculated F and

P values were used to determine the significance of an

input variable. Larger Fvalues and smaller P values are an

indication of the significance of the model. From RSM

analysis, F and P values for the main effects, quadratic

effects and interaction effects of variables were obtained.

The effects of temperatureglucose and induction time

Fig. 2 Surface plot obtained

from optimization using RSM


induction time and glucose on



other parameters constant [21]

Fig. 3 The topology of neural network for the estimation of lipase

production. Triangles represent the inputs (neurons added for ANN

processing); glucose, NaCl, temperature and induction time. Squaresrepresent the hidden and output layer (neurons generated during ANN

processing). Small open circles indicate the input and output layers

(the neurons that can be observed in the form of numerical values)


123


8/12

glucose interactions were significant according to the cal-

culated Fand P values. These values also gave information

about the ranking of variables. For example, a variable with

a large value has a large effect on the response and vice

versa for a variable with a small value.

According to RSM analysis, the maximum effect on

lipase yield was due to temperature. Lipase activity gradu-

ally increased with temperature ranging from 25 to 30C and

then decreased sharply with an increase in temperature from

32 to 45C. The probable reason for the low lipase activity

with respect to high temperature is due to the aggregation of

inclusion bodies. High temperature favours the aggregation

of inclusion bodies whereas lower temperature enhances the

secretion of recombinant protein in soluble form [24]. At

lower temperature, reduced growth of E. coli [2] might be

the reason for decreased lipase production.

The induction time was in second place according to the

significance ranking by RSM. Induction time is one of the

important parameters that affect the expression of recom-

binant proteins. The optimal time of induction (2 h) for

lipase production by recombinant E. coli obtained in this

study corresponded to the early log phase, with cell con-

centrations ranging from 0.3 to 0.5 g/L. Induction was

performed at mid log phase as a base case prior to opti-

mization. The selection of appropriate induction time for

enhancement of recombinant protein production is strain

dependent. The strains, which are sensitive to IPTG con-

centration, need to be induced at late loge phase for

enhancement of recombinant protein production. On the

other hand, strains which are not sensitive to IPTG can be

induced at early log phase [2]. The recombinant E. coli

used in this study for lipase production most probably

belongs to the strain that is not sensitive to IPTG. The

effects of sodium chloride and glucose on lipase production

were not significant and were ranked in third and fourth

place, respectively. RSM analysis also generates informa-

tion on the positive and negative effect that could be

determined by the respective coefficient of the parameters

or variables investigated. The effect of the individual var-

iable or parameter on the process performance could be

evaluated in a more obvious way using RSM as compared

to ANN [8].

Table 3 Sensitivity analysis by ANN for optimization of lipase production by recombinant E. coli

Parameter Glucose (X1) (g/L) NaCl (X2) (g/L) Temperature (X3) (C) Induction time (X4) (h)

Ratio 1.517316 0.960510 3.990739 2.839100

Rank 3.000000 4.000000 1.000000 2.000000

Ratios are values given by ANN as a result of sensitivity analysis to input variables

Rank is the order according to the ratios


from optimization using ANN


temperature and glucose on



other parameters constant


123


9/12

RSM gives quantitative interaction effects of paired

variables, which cannot be estimated by ANN. From RSM,

it was observed that the glucosetemperature and glucoseinduction time interaction effects have P values smaller

than 0.05, indicating that these interaction effects were

significant. The glucosetemperature interaction effect

(F value is 12.08) was stronger than the temperature

induction time interaction effect (F value is 4.83). Other

interaction effects were not significant as they have P val-

ues greater than 0.5.

ANN is widely used in various processes to solve

problems related to prediction, classification and control

[7, 9, 12, 19, 26]. ANN is a modern modelling technique

capable of modelling extremely complex systems, in which

any non-linear systems with large numbers of variables canbe handled. The ANN architecture is represented by its

topology rather than by a model as used in RSM. Topology

in ANN mostly consists of three layers: one input layer,

one hidden layer and one output layer. In most of the

estimation cases one hidden layer is sufficient. Two or

more layers are used for the systems with discontinuities

[4]. The number of hidden neurons is determined according

to the size of the input vector and the number of input

output space classifications. Under-fitting may occur when


from optimization using ANN


induction time and glucose on



other parameters constant

Table 4 Verification experiments including the predicted optimal levels and lipase activity obtained from optimization of lipase production by

recombinant E. coli using RSM and ANN

Serial

no.

Method Glucose

(X1) (g/L)

NaCl (X2)

(g/L)

Temperature

(X3) (C)

Induction

time (X4) (h)

Lipase activity (Y) (IU/mL)

Predicted by

RSM (% diff.a)

Predicted by

ANN (% diff.a)

Experimental

data

1 Base case 50 5 37 3 35.3 (19) 34.7 (17) 29.5 0.799

2 RSM predicted optima 32 5 32 2.12 48.9 (3) 51.3 (2) 50.2 0.809

3 ANN predicted optima 25 3 30 2 39.6 (29) 56.2 (0.7) 55.8 0.934

4 Optima before optimization 10 5 37 2.5 25.2 (13) 30.6 (6) 28.9 0.869

5 Verification experiment 20 3 30 2 38.2 (25) 52.3 (3) 50.6 1.267



Values followed by are standard errors of mean valuesa % difference from the observed values


123


10/12

the number of neurons used in the hidden layer is too small

and vice versa [16]. The developed network in this study

contains three layers: (1) an input layer with four neurons,

(2) one hidden layer with nine neurons and (3) an output

layer with one neuron. The number of hidden neurons is

not too large or too small and represents the best network

system from the observed 60 networks. An increase in the

number of hidden layers and number of neurons therein didnot increase the prediction accuracy.

ANN sensitivity analysis could be carried out to calcu-

late the ratio and thereby the ranking of the variables. Ratio

is the basic measure of sensitivity, which is calculated as

the ratio of the error with missing value substitution to the

original error. Ratio shows how much the network is sen-

sitive for a particular input value. The variable which has a

ratio equal to or less than 1 is not significant, whereas the

variable with a ratio larger than 1 has a significant effect on

output [18]. From the sensitivity analysis conducted in this

study, temperature (ratio = 3.99) gave the highest effect

on lipase production by recombinant E. coli and the lowesteffect was from sodium chloride (ratio = 0.96) (Table 3).

According to significance criteria, the effect of tempera-

ture, induction time and glucose were significant whereas

the effect of NaCl was insignificant. Results from RSM and

ANN indicate that the ranking for temperature and induc-

tion time was similar, whereas the ranking for glucose and

NaCl was different. It is important to note that the qua-

dratic and interaction effects could not be evaluated using

ANN.

The optimum level for glucose (25 g/L) as predicted by

ANN was smaller than that predicted by RSM (32 g/L).

The levels of NaCl (3 g/L), temperature (30C) and

induction time (2 h) predicted by ANN were also lower

than that those predicted by RSM. The optimum induction

time predicted by ANN was similar to that predicted by

RSM. Higher lipase activity was observed at optimum

levels predicted by ANN as compared to RSM. The lipase

activity predicted by ANN also fitted the observed data at

the optimum levels better than the case for RSM (Table 4).

Results from this study showed that the estimation capa-

bility of ANN for optimum levels of variables for lipase

production by recombinant E. coli was more accurate than

RSM. Similar observations have also been reported [8, 28,

29]. Both methods (RSM and ANN) estimated low glucose

concentration as optimal levels. The production of acetate

and other organic acids, as indicated by a drastic drop in

the final culture pH at higher glucose levels, may reduce

the growth and lipase production (Table 1).

Several methods could be used to evaluate the goodness

and accuracy of a given model and to compare two or more

models. The overall predictive capability of the model is

normally determined by R2, but the efficiency of a model

may not be explained by R2 alone [9]. Besides R2, adjusted

R2 may be used to support the models accuracy. In a

multiple linear regression model, adjusted R2 measures the

proportion of the variation in the dependent variable

accounted for by the explanatory variables. Adjusted R2 is

generally considered to be a more accurate goodness-of-fit

measure than R2. Results from this study show that the

adjusted R2 values for both RSM and ANN support the

models accuracy as their values are not much differentfrom R2 values. For a good model, R2 must be closed to 1

whereas the difference between the values of adjusted R2

and R2 must be very small.

Large values of R2 and adjusted R2 do not always mean

that the regression model is an efficient model. Other

values such as AAD and RMSE are also used to validate

and compare more than one model. For a good model, the

AAD value must be as small as possible whereas the

RMSE value must be close to zero. The AAD value (26.38)

for RSM was almost five times higher than the AAD value

(5.09) for ANN (Table 5). In addition, the RMSE value

(1.48) for RSM was two times higher than the RMSE

value (0.63) for ANN. Larger values of RMSE and AAD

mean higher chances of errors in prediction. Therefore,

ANN predictions with lower RMSE and ADD values are

more reliable and accurate than RSM predictions.

The ANN optima were closer to the true optima as

observed from the verification experiments (Table 4). It

was noted that the optimum levels of most of the variables

predicted by RSM and ANN were closer to the edges of

topology. In some other cases as reported in the literature,

the optima predicted by ANN and RSM were located at the

edges [79] and in one case the optima were even located

outside of the topology [8]. It was also observed from the

present studies that the RSM-predicted optima was closer

to the centre points as compared to the ANN-predicted

optima.

The present study indicated that the values predicted by

RSM and ANN were similar to the observed values

towards the centre points. However, the values predicted

by RSM were different from the observed values towards

the edges. At the edges, the effect was observed mainly

from glucose concentration. When glucose concentration

Table 5 Comparison of optimization and prediction capability by

ANN and RSM for lipase production by recombinant E. coli

Serial no. Statistic RSM ANN

1 R2 0.97 0.99

2 Adjusted R2 0.965 0.988

3 AAD 26.38 5.09

4 RMSE 1.48 0.63

AAD absolute average deviation, RMSE root mean square error,

R2 coefficient of correlation determination, adjusted R2 adjusted

coefficient of correlation determination


123


11/12

was decreased from the RSM-predicted optimum glucose

level, the lipase activity predicted by RSM decreased

sharply. Therefore, it can be concluded that RSM-predicted

optima were not as close to the real optima as the ANN-

predicted optima of the lipase production by recombinant

E. coli. For many other bioprocesses, ANN was a better

optima predictor than RSM [79].

Conclusion

Results from this study have demonstrated that ANN gave

better estimation capabilities throughout the range of

variables as compared to RSM in the optimization of lipase

production by a recombinant E. coli. Lipase production

predicted by ANN at optimal variables fitted well to the

experimental data. The constructed ANN has larger R2 and

adjusted R2 values, whereas AAD and RMSE values are

smaller as compared to those observed from the application

of RSM. Similar to the application of RSM, ANN couldalso rank the independent variables. In addition, ANN does

not need a model for prediction, a specified design or

preliminary knowledge about the system. However, the

application of RSM enables the estimation of the quantity

of effect of independent variables as well as the interaction

effects.

Acknowledgments The work presented here is part of Ms Rubina

Nelofers Ph.D. studies, funded by the Pakistan Council of Scientific

and Industrial Research. The research work is funded by Universiti

Putra Malaysia.

References

1. Azaman SN, Ramanan RN, Tan JS, Rahim RA, Abdulla MP,

Ariff AB (2010) Optimization of an induction strategy for

improving interferon-alpha2b production in the periplasm of

Escherichia coli using response surface methodology. Biotechnol

Appl Biochem 56:141150

2. Azaman SN, Ramanan RN, Tan JS, Rahim RA, Abdullah MP,

Ariff AB (2010) Screening for the optimal induction parameters

for periplasmic producing interferon-2b in Escherichia coli. Afr J

Biotechnol 9:63456354

3. Bas D, Boyaci IH (2007) Modeling and optimization I: usability

of response surface methodology. J Food Eng 78:8368454. Basheer IA, Hajmeer M (2000) Artificial neural networks: fun-

damentals, computing, design, and application. J Microbiol

Method 43:331

5. Bezerra MA, Santelli RE, Oliveira EP, Villar SL, Escaleira LA

(2008) Response surface methodology (RSM) as a tool for opti-

mization in analytical chemistry. Talanta 76:965977

6. Cornish-Bowden A (2001) Detection of errors of interpretation in

experiments in enzyme kinetics. Methods 24:181190

7. Dasari VRRK, Donthireddy SRR, Nikku MY, Garapati HR

(2009) Optimization of medium constituents for cephalosporin C

production using response surface methodology and artificial

neural networks. J Biochem Technol 1:6974

8. Desai KM, Survase SA, Saudagar PS, Lele SS, Singhal RS (2008)

Comparison of artificial neural network (ANN) and response

surface methodology (RSM) in fermentation media optimization:

case study of fermentative production of scleroglucan. Biochem

Eng J 41:266273

9. Ebrahimpour A, Rahman RNZRA, Chng DH, Basri M, Salleh

AB (2008) A modelling study by response surface methodology

and artificial neural network on culture parameters optimization

for thermostable lipase production from a newly isolated ther-

mophilic Geobacillus sp. strain ARM. BMC Biotechnol 8:96

10. Eltaweel MA, Rahman RNZRA, Salleh AB, Basri M (2005) An

organic solvent-stable lipase from Bacillus sp. strain 42. Ann

Microbiol 55:187192

11. Farliahati MR, Ramanan NR, Mohamad R, Puspaningsih NNT,

Ariff AB (2010) Enhanced production of xylanase by recombi-

nant Escherichia coli DH5 through optimization of medium

composition using response surface methodology. Ann Microbiol

60:279285

12. Ghaffari A, Abdollahi H, Khoshayand MR, Bozchalooi IS,

Dadgar A, Rafiee-Tehrani M (2006) Performance comparison of

neural network training algorithms in modeling of bimodal drug

delivery. Int J Pharm 327:126138

13. Haider MA, Pakshirajan K, Singh A, Chaudhry S (2008) Artifi-

cial neural network-genetic algorithm approach to optimize

media constituents for enhancing lipase production by a soil

microorganism. Appl Biochem Biotechnol 144:225235

14. Hamid THTA, Eltaweel MA, Rahman RNZRA, Basri M, Salleh

AB (2009) Characterization and solvent stable features of strep-

tagged purified recombinant lipase from thermostable and solvent

tolerant Bacillus sp. strain 42. Ann Microbiol 59:111118

15. Hill T, Lewicki P (2007) Statistics: methods and applications.

StatSoft, Tulsa

16. Karnik SR, Gaitonde VN, Davim JP (2008) A comparative study

of the ANN and RSM modeling approaches for predicting burr

size in drilling. Int J Adv Manuf Technol 38:868883

17. Kwon DY, Rhee JS (1986) A simple and rapid colorimetric

method for determination of free fatty acids for lipase assay.

J Am Oil Chem Soc 63:8992

18. Lou W, Nakai S (2001) Application of artificial neural networks

for predicting the thermal inactivation of bacteria: a combined

effect of temperature, pH and water activity. Food Res Int

34:573579

19. Low CT, Mohamed R, Tan CP, Long K, Ismail R, Lo SK, Lai OM

(2007) Lipase-catalyzed production of medium-chain triacylgly-

cerols from palm kernel oil distillate: optimization using response

surface methodology. Eur J Lipid Sci Technol 109:107119

20. Maldonado LMTP, Hernandez VEB, Rivero EM, Rosa AB,

Flores JLF, Acevedo LGO, Rodrguez ADL (2007) Optimization

of culture conditions for a synthetic gene expression in Esche-

richia coli using response surface methodology: the case of

human interferon beta. Biomolecul Eng 24:217222

21. Nelofer R, Ramanan NR, Rahman RNZRA, Basri M, Ariff AB

(2010) Sequential optimization of production of a thermostable

and organic solvent tolerant lipase by recombinant Escherichiacoli. Ann Microbiol. doi:10.1007/s13213-010-0170-9

22. Noorossana R, Davanloo TS, Saghaei A (2009) An artificial

neural network approach to multiple-response optimization. Int J

Adv Manuf Technol 40:12271238

23. Pan H, Xie Z, Bao W, Zhang J (2008) Optimization of culture

conditions to enhance cis-epoxysuccinate hydrolase production in

Escherichia coli by response surface methodology. Biochem Eng

J 42:133138

24. Qiao CL, Shen BC, Xing JM, Huang J, Zhang JL, Zhao DH,

Yang B (2006) Culture and characteristics of recombinant protein

production of an Escherichia coli strain expressing carboxylest-

erase B1. Int Biodeterior Biodegrad 58:7781


123
http://dx.doi.org/10.1007/s13213-010-0170-9http://dx.doi.org/10.1007/s13213-010-0170-9


12/12

25. Rajendran A, Palanisamy A, Thangavelu V (2008) Evaluation of

medium components by Plackett-Burman statistical design for

lipase production by Candida rugosa and kinetic modelling. Chin

J Biotechnol 24:436444

26. Razmi-Rad E, Ghanbarzadeh B, Rashmekarim J (2008) An arti-

ficial neural network for prediction of Zeleny sedimentation

volume of wheat flour. Int J Agric Biol 10:422426

27. Singh V, Khan M, Khan S, Tripathi CKM (2009) Optimization of

actinomycin V production by Streptomyces triostinicus using

artificial neural network and genetic algorithm. Appl Microbiol

Biotechnol 82:379385

28. Tsao CC (2008) Comparison between response surface method-

ology and radial basis function network for core-center drill in

drilling composite materials. Int J Adv Manuf Technol 37:1061

1068

29. Youssefi SH, Emam-Djomeh Z, Mousavi SM (2009) Comparison

of artificial neural network (ANN) and response surface meth-

odology (RSM) in the prediction of quality parameters of spray-

dried pomegranate juice. Dry Technol 27:910917


123

comparison of the estimation capabilities of response surface

Documents